llama: save more VRAM by reserving n_outputs == n_seqs when possible

[am17an]

Overview

continue #23764, this PR only reserves logits space for n_seqs when possible. With -ub 2048 and MTP, this saves another 1.2GB of VRAM for me. I've tested llama-perplexity also and it seems to work fine. But maybe there is a better API, putting up as a draft for now According to me an API in llama-context is a good solution for this, by default it will reserve all tokens but specifically in server-context we can set it to 1 whenever possible.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, for review and testing across various configs

[drauh]

To motivate this PR further, here is the experiment I ran (copied from the issue #23527 which would be closed by this PR). This patch provides memory reduction even if not using MTP leaving some headroom to put more experts on the GPU for instance.

Setup: Qwen3.6-35B-A3B-UD-Q4_K_XL, -c 150000 -b 4096 -ub 4096 -fa on -ctk q8_0 -ctv q8_0 --kv-unified --cpu-moe -np 1, RTX 3080 10 GB. Peak sampled at 10 Hz over a 145041-token PP.

run	CUDA0 compute	CUDA_Host compute	idle VRAM	peak VRAM
baseline	3976 MiB	2408 MiB	7981 MiB	8275 MiB
#23861 only	3892 MiB	2408 MiB	7897 MiB	8195 MiB
#23764 only	3976 MiB	1236 MiB	7979 MiB	8277 MiB
#23861+ #23764	1982 MiB	1236 MiB	5985 MiB	6279 MiB

[joaotbelo]

Corroborating data on Blackwell (SM_120, RTX 5070 Ti 16 GB) for a config not yet in this thread: MTP draft decoding + vision (mmproj) + --parallel 2, real assistant workload.

Built this branch (810aa71, includes #23764) for CUDA_DOCKER_ARCH=120 and compared against b9426 (already carries #23764), so this isolates the #23861 increment.

Config (identical both runs): Qwen3.6-35B-A3B-UD-Q2_K_XL-MTP, -c 131072 --parallel 2, --spec-type draft-mtp, vision mmproj loaded, -ctk q4_0 -ctv q4_0, -fa on, default -b 2048 -ub 512. Workload: 1 vision request + 2× concurrent ~5.6k-token prefill → 256-tok decode; peak sampled at 5 Hz.

run	idle VRAM	peak VRAM
b9426 (#23764 only)	15495 MiB	15513 MiB
+ #23861	14835 MiB	14859 MiB
saving	660 MiB	654 MiB

Smaller than the -ub 4096 numbers above (expected — the reservation scales with n_ubatch; we run the default -ub 512), but on a 16 GB card it nearly doubles free headroom at peak (790 → 1444 MiB). No OOM, MTP draft acceptance unchanged (~0.5–0.72), and the saving holds at idle and under load — so it's reclaiming a persistently-reserved buffer, not just a prefill spike. 👍